- Published on
Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents
- Authors
- Name
- DeepHorus
Junyang Lin, the former technical lead of Alibaba's Qwen project, who stepped down on March 3, 2026, delivered a talk titled 'Qwen: Towards a Generalist Model / Agent,' ending with a provocative line: 'Training models -> training agents.' As an independent researcher, he later expanded on this vision in a detailed post, critiquing the hybrid thinking approach and advocating for agentic reasoning. The talk and post together signal a major shift in AI development, emphasizing closed-loop interaction over pure reasoning.
Lin's talk toured the Qwen model family, including QwQ-32B, Qwen2.5-Max, Qwen3, and multimodal variants, showing benchmark comparisons against DeepSeek-R1, Grok 3 Beta, Gemini 2.5 Pro, and OpenAI's o-series. The Qwen3 segment stood out with its 'hybrid thinking' feature, combining a thinking mode for step-by-step reasoning and a non-thinking mode for instant responses, alongside dynamic thinking budgets. The model expanded multilingual support from 29 to 119 languages and came in sizes from 0.6B to 235B parameters, all under Apache 2.0 with quantized formats like GGUF and GPTQ.
Despite its apparent smoothness, Lin explained why hybrid thinking was difficult to build. Instruct mode rewards directness, brevity, and low latency, while thinking mode rewards spending more tokens on hard problems. A careless merge degrades both, bloating thinking and dulling instruct performance. Qwen3 attempted a four-stage post-training pipeline, including a long-CoT cold start and a 'thinking mode fusion' step, but later the 2507 line shipped separate Instruct and Thinking variants. Lin framed this as a data problem, not a model problem.
Lin pointed to Anthropic's opposite approach as a useful corrective. Claude 3.7 Sonnet shipped as a hybrid with user-set thinking budgets, and Claude 4 interleaved reasoning with tool use for coding and long tasks. His key insight: a longer reasoning trace does not make a model smarter; thinking should be shaped by the target workload, not benchmarks. This perspective challenges the prevailing reasoning-first paradigm.
Lin drew a clear line between two eras. The first, reasoning thinking, defined by o1 and DeepSeek-R1, taught the field that RL needs deterministic rewards, making math and code central. The next era, agentic thinking, focuses on thinking in order to act. An agent formulates plans, decides when to act, uses tools, reads environmental feedback, and revises after failures—a closed-loop interaction with the world, not an internal monologue.
Agentic thinking must handle decisions that pure reasoning avoids: when to stop thinking and act, which tools to invoke in what order, noisy observations, and plan revision after failures. Lin listed these requirements as critical for future AI. His future work slide for Qwen pointed to more pretraining, RL with environment feedback, longer context, and more modalities, all converging on agents.
As Lin moves from leading Qwen to independent research, his advocacy for agentic thinking marks a pivot from reasoning dominance. The industry may follow, shifting from training static models to training adaptive agents capable of real-world interaction.
